I will present an overview of our work on the automatic, computer-aided design of optical nanostructures. In the past few years, we have developed neural networks that can provide new designs for optical nanostructures with predefined optical response. These neural networks are trained with a large dataset obtained from numerical electromagnetic simulations. In particular, we have developed a conditional generative adversarial network that can predict lithographic masks for meta-atoms with desired transmission and reflection properties. We illustrate our machine-learning-based inverse design with examples of metasurfaces with refractive properties and a Zernike phase mask containing meta-atoms with interdependent properties. Finally, I will discuss some techniques to reduce the amount of simulations needed for the training data set.